[LU-6741] OSP should use inline data for small requests

• Type: Bug
• Resolution: Fixed
• Priority: Critical
• Fix Version/s: Lustre 2.8.0
• Affects Version/s: Lustre 2.8.0
• Labels: dne2
• Severity: 3
• Assignee: Di Wang
• Reporter: Alex Zhuravlev

Description

we shouldn't use bulks to transfer updates unless it's needed. otherwise all "interactive" requests (like lookups, getattrs, etc) need 2 RTTs instead of one.

Issue Links

• is related to: LU-6831 The ticket for tracking all DNE2 bugs
• is related to: LU-7128 Performance improvements for out_read()

Comments

[Peter Jones]

Landed for 2.8

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch http://review.whamcloud.com/15899/
Subject: LU-6741 osp: bulk transfer for osp_md_read
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 28638e84cde481d0216ffc1d170e272bc2cccc76

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch http://review.whamcloud.com/16353/
Subject: LU-6741 osp: Pack small request inline
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: dcaf8a6724ef36f2124428a755b5bb39bcd0c5fd

[Gerrit Updater]

wangdi ([e-mail]) uploaded a new patch: http://review.whamcloud.com/16353
Subject: LU-6741 osp: Pack small request inline
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 94431647ecb69a709928d56bf13104e20843e397

[Gerrit Updater]

wangdi ([e-mail]) uploaded a new patch: http://review.whamcloud.com/15899
Subject: LU-6741 osp: Pack small request inline
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: c1449b11107442b6cfc81287709b64ac83c5f893

[Alex Zhuravlev]

LFSCK uses this protocol to access data on OST, so this will be affected too.

[Oleg Drokin]

So is the perf regression only for DNE, DNE2? Or would single MDT configs be affected as well?

[Di Wang]

This seems important, so I will work on a patch before 2.8 release.

[Alex Zhuravlev]

well, the older versions don't support bulks, right? so, i'd think it's a protocol change. then... I'd say this is a performance regression because in the current release we don't need 2*RTT to fetch attributes (e.g., used all the time by LFSCK), it's a single RTT currently..

[Andreas Dilger]

Alex, what is the impact of this issue? It seems like only a potential performance improvement, but if this is a protocol change for the Async Commit code then I guess it needs to be in 2.8.0?